New Framework L2CU Enhances Human‑AI Classification with Unseen Users

Global: New Framework L2CU Enhances Human‑AI Classification with Unseen Users
A team of machine learning researchers announced a new framework called L2CU on Jan 3, 2026, aiming to improve cooperative classification between humans and AI when the human users have not been seen during training. The work, authored by Dileepa Pitawela, Gustavo Carneiro, and Hsiang‑Ting Chen, was submitted to arXiv and later accepted for publication in IEEE Access.

Problem Addressed

Existing learning‑to‑complement (L2C) methods typically rely on a single, global user model that overlooks individual variability among annotators. Consequently, when a system encounters a new user whose labeling behavior differs from the training set, performance can degrade, limiting the practicality of human‑AI collaboration in real‑world settings.

Core Approach of L2CU

L2CU tackles this limitation by first constructing a set of representative annotator profiles from sparse and noisy user annotations. Each profile captures a distinct labeling pattern. When an unseen user provides a few annotations, L2CU matches the user to the most similar profile and then applies a profile‑specific model to complement the user’s predictions, thereby tailoring assistance to individual behavior.

Evaluation on Diverse Benchmarks

The authors evaluated L2CU on five publicly available datasets: CIFAR‑10N, CIFAR‑10H, Fashion‑MNIST‑H, Chaoyang, and AgNews. Across all benchmarks, L2CU achieved higher joint human‑AI accuracy than baseline L2C methods that use a single user model, demonstrating its effectiveness as a model‑agnostic solution.

Model‑Agnostic Design

Because L2CU operates on top of existing classification models without requiring changes to their architecture, it can be integrated with a wide range of AI systems. This flexibility allows practitioners to adopt the framework in domains ranging from image classification to text categorization.

Implications for Human‑AI Collaboration

By adapting to the unique labeling tendencies of each user, L2CU promises more reliable assistance in settings where expert input is limited or costly. The approach may reduce the amount of training data needed from new users while maintaining high overall performance.

Future Directions

The researchers suggest extending L2CU to other cooperative tasks such as regression and sequence labeling, as well as exploring dynamic profile updates as users evolve over time.
This report is based on information from arXiv, licensed under Academic Preprint / Open Access. Based on the abstract of the research paper. Full text available via ArXiv.